¶1THE ROAD SAYS AIRPORT, SO WE FLY TO PARIS. THAT IS NOT A ROAD. IT'S STOLEN EVIDENCE. GREAT. EUROPE HAS CRIME AND EXCELLENT CANADA.
¶2LET'S CHECK IT. MAN, PARIS, HERE WE COME. AND NOBODY CHECKS FOR LAWYERS. PARIS, HERE WE COME. NOBODY CHECKS.
¶3EVERYONE SLAYS THE LAWYERS. >> THEY ABSOLUTELY CHECK LUGGAGE, PETER. THAT IS NOT A BUSINESS PLAN. I MAR SO THAT PLANE HAS STAIRS, SO IT'S BASICALLY A PETER, BUSES GENERALLY DON'T leave the Exactly. International bus.
¶4>> Okay. So, what you saw in the intro here was kind of my AI infinite stream, right? So, this is something new I really wanted to try out that I saw someone did over the weekend. And they are using kind of the Miniax H3, but they are using kind of a fast version of this. So, this is I think it's like a fine tuned version that runs super fast.
¶5Uh, and what you can achieve with this is actually if we spin up like I kind of down the resolution now to just 480p. But what happens if if we spin this up with like two uh Nvidia B200s on for example run pod uh we can actually generate a video in something like 13 seconds but the clip we generate is 15. Right? So this means that we can kind of run this infinite because for every click we generate, we can play 15 and we can queue up a new one for 13 seconds. So, in our case, we're going to kind of prompt with Luna here from OpenAI to generate like a Family Guy video, and we're going to stream that over on Twitch just using FFmpeg via kind of the run pod with the B200's and that will generate like you will see as you saw in intro like an infinite stream in kind of real time.
¶6Uh but we also have some features as you will see that the users can do some kind of exclamation mark prompt and they can kind of affect where the story goes and that is kind of the whole point of this. So the users that watches can kind of in uh like interact with the story and kind of point the story in right direction. Uh unfortunately we don't have like uh video like image to video yet because that would make it even more coherent but maybe in the future we do something with that. Now it's just text to video. But yeah, uh I just want to go through how I set this up.
¶7So it's going to be some voice over and stuff and you can kind of see how I got this all set up in somewhat of detail, not perfect. But yeah, hope you enjoy this and I'll come back at the end and talk a bit more about uh some few other stuff. So the framework we are kind of using for this is uh something called infinite live streams. This kind of connects to your Twitch and we can use like yeah input that kind of affects the setup and we're using the H3 fast. This is I think it's like a fine tune of the Miniax H3 that works superb for kind of this use case.
¶8So basically what I did now is just fired up Codex. I let my SER API agent loop do the research, find everything we needed to know about this pipeline and from there basically Codex could handle everything. And that was kind of perfect because yeah, today's sponsor is SER API. So today's video is sponsored by SER API. SER API, they provide clean, real-time structured search results data from Google, YouTube, and more as you will see me using soon.
¶9They handle kind of the frustrating parts of web scraping like solving captions, rotating proxies, adapting to changing layouts, so developers can focus on building. that makes it especially useful for AI agents like we are doing today which kind of needs current information in like a reliable JSON format output. Okay, so today is earnings call for Nvidia and I'm about to build an AI agent that is going to look at this Cali earnings market that we probably going to set up some AI agentic trading on. But to do that, I need to gather a lot of context because if you're going to look at Google here, all of this uh Nvidia news around the earnings changes all the time. So today, I'm going to actually use the sponsor SER API and I really enjoy kind of their structured way of uh collecting data uh by yeah, for example, you can do it via the Google search API.
¶10Uh they also have an API for YouTube search. So, we're going to try to combo uh combine both and actually set up our AI agent with the Google search for Nvidia earnings and the YouTube search. And let's just see what kind of results we can get from that. So, now I just headed over to Codeex. We are setting up an AI agent to get context for the earnings call today from Nvidia.
¶11We are going to use the Google search API, YouTube search API. Uh my key is in my ENV and write the script for this test then execute. Uh use different search terms. We can specify that and bring us back the following sentiment around earnings call, expected surprise percentage and information we can find actually about the what will Nvidia say during the next earnings call just to test it out. Bring this in a structured uh JSON format because we know that SER API has that.
¶12Here you can kind of see in the results now why I really like uh these results from SER API. You can see we have the event, company, ticker, Nvidia, right? Sentiment around earnings calls positive. So we have 30 positive hits, nine negative. So I just say uh write all of this out in bullet points with links to sources.
¶13For example, you can see we have uh yeah, we have some YouTube videos here. We can open this. I guess this hasn't started. This is like a live stream coming in today. So this is why I really like using SER API to kind of set up my AI agents with uh the kind of tools to do uh research on its own and not just rely on for example this case codeex web search because we get so much more if we're using SER API.
¶14So get started with SER API using 250 create credits. You can get that by clicking the link in the description or just scan the QR code that you see on the screen here. Okay. So now we are back on Codeex. We have started this up and here you can see my run pub.
¶15These are the two BT200s I have kind of rented here. And you can see we have filled this up now. 155 GB on the yeah volume I guess. And here you can see now we are kind of firing up the kernel that completed. Now we are installing some Q.
¶16PyTorch packages onto the to the pod. And yeah, we're just going to keep this up and you can see it's just going to be like a yeah small cost here to stage the GPU. And of course we need to prepare our Twitch here. So everything is just my yeah basically just my old channel and it shouldn't be too hard to kind of prepare this Twitch here. And now you can see we have built our environment that is kind of including the CUDA, the 13 PyTorch and of course the B200 kernel.
¶17And yeah, now we should be able to kind of fire up the GPUs because we loaded all the model weights into the system. And you can see you can see here now the price is uh 13 something per hour and we don't have too much time. So you can see our Twitch stream is ready. That's pretty good. And now we're probably just going to wait for the weight to load and we need to load up kind of the full 150 GB something model into the memory of the GPU.
¶18And you can see we are starting this now I think uh loading up the the model weights into the GPUs. Uh and yeah shouldn't take cost too much now. But as long as the the GPUs are running of course we need to pay that premium $14 or something per hour. So it kind of helps to line everything up before you kind of spin up the GPUs. But at this stage we just have to spin them up because we need to load the weights.
¶19And hopefully now you can see uh we have allocated this at 1350 per hour and I think we should be actually ready now if I remember correctly. Uh yeah, I think I just loaded everything up. I started it again and we're going to test it. I think I did here just to generate a clip and hopefully here now you can see uh yeah, we have $26. We have 1 hour and 54 minutes left to run.
¶20But now we should have loaded all the the weights. Yeah. And our GPUs is actually spinning here now. V RAM. Yeah, we loaded up 132 GB of memory and the GPUs have started.
¶21So now we are kind of ready actually to do this and we should spin up to 100 now. Yeah. Yeah. Perfect. So the logic is basically if no one is putting in anything in the chat, we have the Luna to generate auto clips.
¶22That kind of builds the story autonomous. Uh but if someone puts in like exclamation mark prompt and that's going to be prioritized over the auto cue for the story. So we can see yes use this as a suggestion. If not autonom autonomously continue the story by using open luna and yeah we're just going to validate this. We're going to send the prompt to the fast h3 model.
¶23Uh yeah, we're going to fast three. It's going to render this super fast on the B200's and we're going to add it add this to the queue and we're going to save the story to memory, play it on Twitch and then we go back to the queue again so we can kind of do some kind of coherent story. So we always try to keep it somewhat coherent, right? That is kind of the point of this. Uh now you can see we should be pretty much ready now to load everything up and start um the Twitch stream.
¶24So all of this is like a pipeline and if everything kind of succeed it should just load up those eight clips start playing back to Twitch and we should be able to kind of see something happening on the screen now. So everything is corrected. We generated the clip here in like 10 seconds. So that's pretty good because of the 15-second clip. And that means that we can kind of keep this everything.
¶25So let's just watch a couple of clips now and see if everything works. >> Great. I'm finally GETTING SHIPPED SOMEWHERE. BRIAN, TELL HIM I'M NOT LUGGAGE. BRIAN IS NOT AFRAID.
¶26LET ME BAD. I SAID PRIAN. PRIOR THINK I'M NOT LUGGAGE. YOUR BADGE JUST APPROVED YOUR GREAT. I finally have a career with benefits.
¶27TWICE. TWICE. I WILL FAVOR IT. I'm a trip wicked her. BRIAN, STOP SHIPPING ME TO MYSTERY COUNTRIES.
¶28I'm changing the destination to not my great. That's where my taxes go. Help spin collite hit. >> Bayan, TELL A BOX I OUTRANK GRAVITY. >> This car that goes LIKE AN EPISODE.
¶29>> THAT BADGE SAYS CARGO, not genius. Then promote me before I hit the FBI. >> Okay, so you can see this is working pretty good, y'all. You can see up in the right corner there says ready. These are all the clips that are kind of queued up and building is when uh either OpenAI is running the Luna model to build up a new scene or if anyone is saying in the chat like exclamation mark prompt and is trying to interfere with the story and you can see kind of we have a Q system here what is coming up next.
¶30So now let's try to type in something here. So I'm going to say prompt Peter and Brian gets on a plane to Europe and Paris quit the startup and start enjoying life. And now we're going to send this into the chat. And that means that we should trigger a new story type, right? So you can kind of see now this is already lined up.
¶31So let's just listen how this turned out. Now >> shoot over the head towards us to lift. THE ROAD SAYS AIRPORT. SO WE FLY TO PARIS. THAT IS NOT A ROAD.
¶32IT'S STOLEN EVIDENCE. >> GREAT. EUROPE HAS CRIME AND EXCELLENT CANADA. LET'S CHECK IT. MAN, PARIS, HERE WE COME.
¶33AND NOBODY CHECKS LAWYERS. PARIS, HERE WE COME. NOBODY CHECKS. EVERYONE SLAYS THE LAWYERS. >> THEY ABSOLUTELY CHECK LUGGAGE.
¶34PETER, THAT IS NOT A BUSINESS PLAN. I SHOULD tell us. THAT PLANE HAS STAIRS, so it's basically A PETER. BUSES GENERALLY don't leave the Exactly. international bus.
¶35>> Okay, so that worked out pretty good. Now let's try Peter screams uh in French wee wee or something and let's see how that turns out. >> Good investors love mobility. Dig my guess FOR I WON'T EVER MAN. WEE WEE.
¶36GET THE LAPTOP FANCY DRAIN MAN. PETER YOU'RE CHEERING THE DISASTER. >> I WON'T BE AT HAMMING H. It's French for excellent business strategy. >> Another thing we can do is we can change the direction.
¶37So we can do dash movie. Peter uh gets a role in a big Hollywood movie. This kind of changes the whole scene here. >> I'm going to be in the biggest movie ever. >> You still need to audition, Peter.
¶38>> Great. I'll audition for the chicken sandwich. I'M SHOWING SIGHT. I ORDERED EVERY DAY. I am ready for my Hollywood food audition.
¶39I sn I have to venture the fashion. >> For $20, I can teach you acting. >> That toast you THAT THOUGH ALREADY MADE him cast his rain that way. Then coach me how TO DRIVE THERE. A HASH.
¶40I WENT OVER AND THAT WAS THE SMACK THINGY. All right, JULIUS, START. I'LL OPEN IT WITH STAR POWER. >> That is not how mailboxes work. >> Then Hollywood needs better mailboxes >> where I am now.
¶41So, I got to say I really enjoyed this experiment and it's going to be interesting to see if this is going to catch on. But kind of the big thing now is kind of the $14 price point, right? It's really it's really kind of hard to overcome that at the moment because then you kind of need like if you had like kind of like a business model here where you can kind of collect some money and you kind of can overcome the the price point. Uh, but actually if you want to run it in like 720p, I think you need eight. You need eight B200s and that's going to rack up like $115 or something.
¶42$100 at least an hour. So, you will need kind of a lot of financing if you're going to run this 24/7. And you probably will get banned here on Twitch to if you kind of stream this 24/7. Uh, but it is interesting, right? And I thought it was pretty cool to just go through this, set it up, and learn more about it because I think personally this is going to be something that is going to be more prevalent in the future.
¶43Uh, but we just have to see how much compute that is actually available. So yeah, really fun and I really enjoy checking out this H3 fast setup with the setup we have here. So uh yeah, uh don't forget to check out SER API, today's sponsor. You will find the link in the description below if you want to do research on things like I did in this video to set everything up here. Uh worked super well and you can do it to build AI agents and stuff like that like I did in the showcase.
¶44So yeah, hope you enjoy this kind of new different thing I wanted to just do a video on and yeah, hopefully I'll see you again soon. So, have a nice day.